ManaTEE: Enabling Verifiable AI Transparency With Confidential Computing
summary
ManaTEE is an open-source ManaTEE framework that runs model evaluation inside attested TEEs, exposing cryptographic evidence of what code and data ran. It uses a secure Jupyter interface and the LLM Evaluation Harness to run benchmarks (BBQ, CrowS-Pairs, Simple Co-occurrence, pronoun resolution) and produce reproducible, auditable reports plus an attestation signature. Demo shows submit → execute in a CPU-only TEE → retrieve results and an attestation report; backend is single-node today, GPU support planned.
AI-written summary · about
chapters (9)
- 0:00ManaTEE and confidential computing
- 3:05Data in use needs protection
- 7:00Trusted execution environments
- 10:51Private data analytics challenges
- 15:25ManaTEE framework goals
- 17:54Two-stage execution model
- 21:12Verifying LLM fairness
- 26:29Lessons learned and future
- 27:58Q&A
related talks
36mSecurity for AI and Multi-Party Collaboration with Confidential Computing (Oh, and...
30mProtecting Wasm Workloads with Confidential Computing and TEE...
37mConfidential AI: Enabling Trustworthy Ecosystems in the Age of Agentic Intelligence - Session
19mPrivate AI: Building Trust Through Verifiable Computation
44mGPU-less, Trust-less, Limit-less: Reimagining the Confidential AI Cloud
34mTeaclave: Making Computation on Privacy-Sensitive Data Safe and Simple
42mSponsored Session: Confidential Computing - New Capabilities for New Workloads
40mDoing the Safety Dance: Will Trusted Execution Environments Protect Us?